4 #include "util/evsel.h"
5 #include "util/evlist.h"
7 #include "util/cache.h"
8 #include "util/symbol.h"
9 #include "util/thread.h"
10 #include "util/header.h"
11 #include "util/session.h"
12 #include "util/intlist.h"
13 #include <subcmd/parse-options.h>
14 #include "util/trace-event.h"
15 #include "util/debug.h"
16 #include "util/tool.h"
17 #include "util/stat.h"
19 #include "util/data.h"
20 #include "util/ordered-events.h"
22 #include <sys/prctl.h>
23 #ifdef HAVE_TIMERFD_SUPPORT
24 #include <sys/timerfd.h>
27 #include <linux/time64.h>
29 #include <semaphore.h>
33 #ifdef HAVE_KVM_STAT_SUPPORT
34 #include "util/kvm-stat.h"
36 void exit_event_get_key(struct perf_evsel
*evsel
,
37 struct perf_sample
*sample
,
38 struct event_key
*key
)
41 key
->key
= perf_evsel__intval(evsel
, sample
, kvm_exit_reason
);
44 bool kvm_exit_event(struct perf_evsel
*evsel
)
46 return !strcmp(evsel
->name
, kvm_exit_trace
);
49 bool exit_event_begin(struct perf_evsel
*evsel
,
50 struct perf_sample
*sample
, struct event_key
*key
)
52 if (kvm_exit_event(evsel
)) {
53 exit_event_get_key(evsel
, sample
, key
);
60 bool kvm_entry_event(struct perf_evsel
*evsel
)
62 return !strcmp(evsel
->name
, kvm_entry_trace
);
65 bool exit_event_end(struct perf_evsel
*evsel
,
66 struct perf_sample
*sample __maybe_unused
,
67 struct event_key
*key __maybe_unused
)
69 return kvm_entry_event(evsel
);
72 static const char *get_exit_reason(struct perf_kvm_stat
*kvm
,
73 struct exit_reasons_table
*tbl
,
76 while (tbl
->reason
!= NULL
) {
77 if (tbl
->exit_code
== exit_code
)
82 pr_err("unknown kvm exit code:%lld on %s\n",
83 (unsigned long long)exit_code
, kvm
->exit_reasons_isa
);
87 void exit_event_decode_key(struct perf_kvm_stat
*kvm
,
88 struct event_key
*key
,
91 const char *exit_reason
= get_exit_reason(kvm
, key
->exit_reasons
,
94 scnprintf(decode
, decode_str_len
, "%s", exit_reason
);
97 static bool register_kvm_events_ops(struct perf_kvm_stat
*kvm
)
99 struct kvm_reg_events_ops
*events_ops
= kvm_reg_events_ops
;
101 for (events_ops
= kvm_reg_events_ops
; events_ops
->name
; events_ops
++) {
102 if (!strcmp(events_ops
->name
, kvm
->report_event
)) {
103 kvm
->events_ops
= events_ops
->ops
;
111 struct vcpu_event_record
{
114 struct kvm_event
*last_event
;
118 static void init_kvm_event_record(struct perf_kvm_stat
*kvm
)
122 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++)
123 INIT_LIST_HEAD(&kvm
->kvm_events_cache
[i
]);
126 #ifdef HAVE_TIMERFD_SUPPORT
127 static void clear_events_cache_stats(struct list_head
*kvm_events_cache
)
129 struct list_head
*head
;
130 struct kvm_event
*event
;
134 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
135 head
= &kvm_events_cache
[i
];
136 list_for_each_entry(event
, head
, hash_entry
) {
137 /* reset stats for event */
138 event
->total
.time
= 0;
139 init_stats(&event
->total
.stats
);
141 for (j
= 0; j
< event
->max_vcpu
; ++j
) {
142 event
->vcpu
[j
].time
= 0;
143 init_stats(&event
->vcpu
[j
].stats
);
150 static int kvm_events_hash_fn(u64 key
)
152 return key
& (EVENTS_CACHE_SIZE
- 1);
155 static bool kvm_event_expand(struct kvm_event
*event
, int vcpu_id
)
157 int old_max_vcpu
= event
->max_vcpu
;
160 if (vcpu_id
< event
->max_vcpu
)
163 while (event
->max_vcpu
<= vcpu_id
)
164 event
->max_vcpu
+= DEFAULT_VCPU_NUM
;
167 event
->vcpu
= realloc(event
->vcpu
,
168 event
->max_vcpu
* sizeof(*event
->vcpu
));
171 pr_err("Not enough memory\n");
175 memset(event
->vcpu
+ old_max_vcpu
, 0,
176 (event
->max_vcpu
- old_max_vcpu
) * sizeof(*event
->vcpu
));
180 static struct kvm_event
*kvm_alloc_init_event(struct event_key
*key
)
182 struct kvm_event
*event
;
184 event
= zalloc(sizeof(*event
));
186 pr_err("Not enough memory\n");
191 init_stats(&event
->total
.stats
);
195 static struct kvm_event
*find_create_kvm_event(struct perf_kvm_stat
*kvm
,
196 struct event_key
*key
)
198 struct kvm_event
*event
;
199 struct list_head
*head
;
201 BUG_ON(key
->key
== INVALID_KEY
);
203 head
= &kvm
->kvm_events_cache
[kvm_events_hash_fn(key
->key
)];
204 list_for_each_entry(event
, head
, hash_entry
) {
205 if (event
->key
.key
== key
->key
&& event
->key
.info
== key
->info
)
209 event
= kvm_alloc_init_event(key
);
213 list_add(&event
->hash_entry
, head
);
217 static bool handle_begin_event(struct perf_kvm_stat
*kvm
,
218 struct vcpu_event_record
*vcpu_record
,
219 struct event_key
*key
, u64 timestamp
)
221 struct kvm_event
*event
= NULL
;
223 if (key
->key
!= INVALID_KEY
)
224 event
= find_create_kvm_event(kvm
, key
);
226 vcpu_record
->last_event
= event
;
227 vcpu_record
->start_time
= timestamp
;
232 kvm_update_event_stats(struct kvm_event_stats
*kvm_stats
, u64 time_diff
)
234 kvm_stats
->time
+= time_diff
;
235 update_stats(&kvm_stats
->stats
, time_diff
);
238 static double kvm_event_rel_stddev(int vcpu_id
, struct kvm_event
*event
)
240 struct kvm_event_stats
*kvm_stats
= &event
->total
;
243 kvm_stats
= &event
->vcpu
[vcpu_id
];
245 return rel_stddev_stats(stddev_stats(&kvm_stats
->stats
),
246 avg_stats(&kvm_stats
->stats
));
249 static bool update_kvm_event(struct kvm_event
*event
, int vcpu_id
,
253 kvm_update_event_stats(&event
->total
, time_diff
);
257 if (!kvm_event_expand(event
, vcpu_id
))
260 kvm_update_event_stats(&event
->vcpu
[vcpu_id
], time_diff
);
264 static bool is_child_event(struct perf_kvm_stat
*kvm
,
265 struct perf_evsel
*evsel
,
266 struct perf_sample
*sample
,
267 struct event_key
*key
)
269 struct child_event_ops
*child_ops
;
271 child_ops
= kvm
->events_ops
->child_ops
;
276 for (; child_ops
->name
; child_ops
++) {
277 if (!strcmp(evsel
->name
, child_ops
->name
)) {
278 child_ops
->get_key(evsel
, sample
, key
);
286 static bool handle_child_event(struct perf_kvm_stat
*kvm
,
287 struct vcpu_event_record
*vcpu_record
,
288 struct event_key
*key
,
289 struct perf_sample
*sample __maybe_unused
)
291 struct kvm_event
*event
= NULL
;
293 if (key
->key
!= INVALID_KEY
)
294 event
= find_create_kvm_event(kvm
, key
);
296 vcpu_record
->last_event
= event
;
301 static bool skip_event(const char *event
)
303 const char * const *skip_events
;
305 for (skip_events
= kvm_skip_events
; *skip_events
; skip_events
++)
306 if (!strcmp(event
, *skip_events
))
312 static bool handle_end_event(struct perf_kvm_stat
*kvm
,
313 struct vcpu_event_record
*vcpu_record
,
314 struct event_key
*key
,
315 struct perf_sample
*sample
)
317 struct kvm_event
*event
;
318 u64 time_begin
, time_diff
;
321 if (kvm
->trace_vcpu
== -1)
324 vcpu
= vcpu_record
->vcpu_id
;
326 event
= vcpu_record
->last_event
;
327 time_begin
= vcpu_record
->start_time
;
329 /* The begin event is not caught. */
334 * In some case, the 'begin event' only records the start timestamp,
335 * the actual event is recognized in the 'end event' (e.g. mmio-event).
338 /* Both begin and end events did not get the key. */
339 if (!event
&& key
->key
== INVALID_KEY
)
343 event
= find_create_kvm_event(kvm
, key
);
348 vcpu_record
->last_event
= NULL
;
349 vcpu_record
->start_time
= 0;
351 /* seems to happen once in a while during live mode */
352 if (sample
->time
< time_begin
) {
353 pr_debug("End time before begin time; skipping event.\n");
357 time_diff
= sample
->time
- time_begin
;
359 if (kvm
->duration
&& time_diff
> kvm
->duration
) {
360 char decode
[decode_str_len
];
362 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
363 if (!skip_event(decode
)) {
364 pr_info("%" PRIu64
" VM %d, vcpu %d: %s event took %" PRIu64
"usec\n",
365 sample
->time
, sample
->pid
, vcpu_record
->vcpu_id
,
366 decode
, time_diff
/ NSEC_PER_USEC
);
370 return update_kvm_event(event
, vcpu
, time_diff
);
374 struct vcpu_event_record
*per_vcpu_record(struct thread
*thread
,
375 struct perf_evsel
*evsel
,
376 struct perf_sample
*sample
)
378 /* Only kvm_entry records vcpu id. */
379 if (!thread__priv(thread
) && kvm_entry_event(evsel
)) {
380 struct vcpu_event_record
*vcpu_record
;
382 vcpu_record
= zalloc(sizeof(*vcpu_record
));
384 pr_err("%s: Not enough memory\n", __func__
);
388 vcpu_record
->vcpu_id
= perf_evsel__intval(evsel
, sample
,
390 thread__set_priv(thread
, vcpu_record
);
393 return thread__priv(thread
);
396 static bool handle_kvm_event(struct perf_kvm_stat
*kvm
,
397 struct thread
*thread
,
398 struct perf_evsel
*evsel
,
399 struct perf_sample
*sample
)
401 struct vcpu_event_record
*vcpu_record
;
402 struct event_key key
= { .key
= INVALID_KEY
,
403 .exit_reasons
= kvm
->exit_reasons
};
405 vcpu_record
= per_vcpu_record(thread
, evsel
, sample
);
409 /* only process events for vcpus user cares about */
410 if ((kvm
->trace_vcpu
!= -1) &&
411 (kvm
->trace_vcpu
!= vcpu_record
->vcpu_id
))
414 if (kvm
->events_ops
->is_begin_event(evsel
, sample
, &key
))
415 return handle_begin_event(kvm
, vcpu_record
, &key
, sample
->time
);
417 if (is_child_event(kvm
, evsel
, sample
, &key
))
418 return handle_child_event(kvm
, vcpu_record
, &key
, sample
);
420 if (kvm
->events_ops
->is_end_event(evsel
, sample
, &key
))
421 return handle_end_event(kvm
, vcpu_record
, &key
, sample
);
426 #define GET_EVENT_KEY(func, field) \
427 static u64 get_event_ ##func(struct kvm_event *event, int vcpu) \
430 return event->total.field; \
432 if (vcpu >= event->max_vcpu) \
435 return event->vcpu[vcpu].field; \
438 #define COMPARE_EVENT_KEY(func, field) \
439 GET_EVENT_KEY(func, field) \
440 static int compare_kvm_event_ ## func(struct kvm_event *one, \
441 struct kvm_event *two, int vcpu)\
443 return get_event_ ##func(one, vcpu) > \
444 get_event_ ##func(two, vcpu); \
447 GET_EVENT_KEY(time
, time
);
448 COMPARE_EVENT_KEY(count
, stats
.n
);
449 COMPARE_EVENT_KEY(mean
, stats
.mean
);
450 GET_EVENT_KEY(max
, stats
.max
);
451 GET_EVENT_KEY(min
, stats
.min
);
453 #define DEF_SORT_NAME_KEY(name, compare_key) \
454 { #name, compare_kvm_event_ ## compare_key }
456 static struct kvm_event_key keys
[] = {
457 DEF_SORT_NAME_KEY(sample
, count
),
458 DEF_SORT_NAME_KEY(time
, mean
),
462 static bool select_key(struct perf_kvm_stat
*kvm
)
466 for (i
= 0; keys
[i
].name
; i
++) {
467 if (!strcmp(keys
[i
].name
, kvm
->sort_key
)) {
468 kvm
->compare
= keys
[i
].key
;
473 pr_err("Unknown compare key:%s\n", kvm
->sort_key
);
477 static void insert_to_result(struct rb_root
*result
, struct kvm_event
*event
,
478 key_cmp_fun bigger
, int vcpu
)
480 struct rb_node
**rb
= &result
->rb_node
;
481 struct rb_node
*parent
= NULL
;
485 p
= container_of(*rb
, struct kvm_event
, rb
);
488 if (bigger(event
, p
, vcpu
))
489 rb
= &(*rb
)->rb_left
;
491 rb
= &(*rb
)->rb_right
;
494 rb_link_node(&event
->rb
, parent
, rb
);
495 rb_insert_color(&event
->rb
, result
);
499 update_total_count(struct perf_kvm_stat
*kvm
, struct kvm_event
*event
)
501 int vcpu
= kvm
->trace_vcpu
;
503 kvm
->total_count
+= get_event_count(event
, vcpu
);
504 kvm
->total_time
+= get_event_time(event
, vcpu
);
507 static bool event_is_valid(struct kvm_event
*event
, int vcpu
)
509 return !!get_event_count(event
, vcpu
);
512 static void sort_result(struct perf_kvm_stat
*kvm
)
515 int vcpu
= kvm
->trace_vcpu
;
516 struct kvm_event
*event
;
518 for (i
= 0; i
< EVENTS_CACHE_SIZE
; i
++) {
519 list_for_each_entry(event
, &kvm
->kvm_events_cache
[i
], hash_entry
) {
520 if (event_is_valid(event
, vcpu
)) {
521 update_total_count(kvm
, event
);
522 insert_to_result(&kvm
->result
, event
,
529 /* returns left most element of result, and erase it */
530 static struct kvm_event
*pop_from_result(struct rb_root
*result
)
532 struct rb_node
*node
= rb_first(result
);
537 rb_erase(node
, result
);
538 return container_of(node
, struct kvm_event
, rb
);
541 static void print_vcpu_info(struct perf_kvm_stat
*kvm
)
543 int vcpu
= kvm
->trace_vcpu
;
545 pr_info("Analyze events for ");
547 if (kvm
->opts
.target
.system_wide
)
548 pr_info("all VMs, ");
549 else if (kvm
->opts
.target
.pid
)
550 pr_info("pid(s) %s, ", kvm
->opts
.target
.pid
);
552 pr_info("dazed and confused on what is monitored, ");
555 pr_info("all VCPUs:\n\n");
557 pr_info("VCPU %d:\n\n", vcpu
);
560 static void show_timeofday(void)
566 gettimeofday(&tv
, NULL
);
567 if (localtime_r(&tv
.tv_sec
, <ime
)) {
568 strftime(date
, sizeof(date
), "%H:%M:%S", <ime
);
569 pr_info("%s.%06ld", date
, tv
.tv_usec
);
571 pr_info("00:00:00.000000");
576 static void print_result(struct perf_kvm_stat
*kvm
)
578 char decode
[decode_str_len
];
579 struct kvm_event
*event
;
580 int vcpu
= kvm
->trace_vcpu
;
588 print_vcpu_info(kvm
);
589 pr_info("%*s ", decode_str_len
, kvm
->events_ops
->name
);
590 pr_info("%10s ", "Samples");
591 pr_info("%9s ", "Samples%");
593 pr_info("%9s ", "Time%");
594 pr_info("%11s ", "Min Time");
595 pr_info("%11s ", "Max Time");
596 pr_info("%16s ", "Avg time");
599 while ((event
= pop_from_result(&kvm
->result
))) {
600 u64 ecount
, etime
, max
, min
;
602 ecount
= get_event_count(event
, vcpu
);
603 etime
= get_event_time(event
, vcpu
);
604 max
= get_event_max(event
, vcpu
);
605 min
= get_event_min(event
, vcpu
);
607 kvm
->events_ops
->decode_key(kvm
, &event
->key
, decode
);
608 pr_info("%*s ", decode_str_len
, decode
);
609 pr_info("%10llu ", (unsigned long long)ecount
);
610 pr_info("%8.2f%% ", (double)ecount
/ kvm
->total_count
* 100);
611 pr_info("%8.2f%% ", (double)etime
/ kvm
->total_time
* 100);
612 pr_info("%9.2fus ", (double)min
/ NSEC_PER_USEC
);
613 pr_info("%9.2fus ", (double)max
/ NSEC_PER_USEC
);
614 pr_info("%9.2fus ( +-%7.2f%% )", (double)etime
/ ecount
/ NSEC_PER_USEC
,
615 kvm_event_rel_stddev(vcpu
, event
));
619 pr_info("\nTotal Samples:%" PRIu64
", Total events handled time:%.2fus.\n\n",
620 kvm
->total_count
, kvm
->total_time
/ (double)NSEC_PER_USEC
);
622 if (kvm
->lost_events
)
623 pr_info("\nLost events: %" PRIu64
"\n\n", kvm
->lost_events
);
626 #ifdef HAVE_TIMERFD_SUPPORT
627 static int process_lost_event(struct perf_tool
*tool
,
628 union perf_event
*event __maybe_unused
,
629 struct perf_sample
*sample __maybe_unused
,
630 struct machine
*machine __maybe_unused
)
632 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
, tool
);
639 static bool skip_sample(struct perf_kvm_stat
*kvm
,
640 struct perf_sample
*sample
)
642 if (kvm
->pid_list
&& intlist__find(kvm
->pid_list
, sample
->pid
) == NULL
)
648 static int process_sample_event(struct perf_tool
*tool
,
649 union perf_event
*event
,
650 struct perf_sample
*sample
,
651 struct perf_evsel
*evsel
,
652 struct machine
*machine
)
655 struct thread
*thread
;
656 struct perf_kvm_stat
*kvm
= container_of(tool
, struct perf_kvm_stat
,
659 if (skip_sample(kvm
, sample
))
662 thread
= machine__findnew_thread(machine
, sample
->pid
, sample
->tid
);
663 if (thread
== NULL
) {
664 pr_debug("problem processing %d event, skipping it.\n",
669 if (!handle_kvm_event(kvm
, thread
, evsel
, sample
))
676 static int cpu_isa_config(struct perf_kvm_stat
*kvm
)
678 char buf
[64], *cpuid
;
682 err
= get_cpuid(buf
, sizeof(buf
));
684 pr_err("Failed to look up CPU type\n");
689 cpuid
= kvm
->session
->header
.env
.cpuid
;
692 pr_err("Failed to look up CPU type\n");
696 err
= cpu_isa_init(kvm
, cpuid
);
698 pr_err("CPU %s is not supported.\n", cpuid
);
703 static bool verify_vcpu(int vcpu
)
705 if (vcpu
!= -1 && vcpu
< 0) {
706 pr_err("Invalid vcpu:%d.\n", vcpu
);
713 #ifdef HAVE_TIMERFD_SUPPORT
714 /* keeping the max events to a modest level to keep
715 * the processing of samples per mmap smooth.
717 #define PERF_KVM__MAX_EVENTS_PER_MMAP 25
719 static s64
perf_kvm__mmap_read_idx(struct perf_kvm_stat
*kvm
, int idx
,
722 union perf_event
*event
;
723 struct perf_sample sample
;
727 *mmap_time
= ULLONG_MAX
;
728 while ((event
= perf_evlist__mmap_read(kvm
->evlist
, idx
)) != NULL
) {
729 err
= perf_evlist__parse_sample(kvm
->evlist
, event
, &sample
);
731 perf_evlist__mmap_consume(kvm
->evlist
, idx
);
732 pr_err("Failed to parse sample\n");
736 err
= perf_session__queue_event(kvm
->session
, event
, &sample
, 0);
738 * FIXME: Here we can't consume the event, as perf_session__queue_event will
739 * point to it, and it'll get possibly overwritten by the kernel.
741 perf_evlist__mmap_consume(kvm
->evlist
, idx
);
744 pr_err("Failed to enqueue sample: %d\n", err
);
748 /* save time stamp of our first sample for this mmap */
750 *mmap_time
= sample
.time
;
752 /* limit events per mmap handled all at once */
754 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
761 static int perf_kvm__mmap_read(struct perf_kvm_stat
*kvm
)
763 int i
, err
, throttled
= 0;
765 u64 flush_time
= ULLONG_MAX
, mmap_time
;
767 for (i
= 0; i
< kvm
->evlist
->nr_mmaps
; i
++) {
768 n
= perf_kvm__mmap_read_idx(kvm
, i
, &mmap_time
);
772 /* flush time is going to be the minimum of all the individual
773 * mmap times. Essentially, we flush all the samples queued up
774 * from the last pass under our minimal start time -- that leaves
775 * a very small race for samples to come in with a lower timestamp.
776 * The ioctl to return the perf_clock timestamp should close the
779 if (mmap_time
< flush_time
)
780 flush_time
= mmap_time
;
783 if (n
== PERF_KVM__MAX_EVENTS_PER_MMAP
)
787 /* flush queue after each round in which we processed events */
789 struct ordered_events
*oe
= &kvm
->session
->ordered_events
;
791 oe
->next_flush
= flush_time
;
792 err
= ordered_events__flush(oe
, OE_FLUSH__ROUND
);
794 if (kvm
->lost_events
)
795 pr_info("\nLost events: %" PRIu64
"\n\n",
804 static volatile int done
;
806 static void sig_handler(int sig __maybe_unused
)
811 static int perf_kvm__timerfd_create(struct perf_kvm_stat
*kvm
)
813 struct itimerspec new_value
;
816 kvm
->timerfd
= timerfd_create(CLOCK_MONOTONIC
, TFD_NONBLOCK
);
817 if (kvm
->timerfd
< 0) {
818 pr_err("timerfd_create failed\n");
822 new_value
.it_value
.tv_sec
= kvm
->display_time
;
823 new_value
.it_value
.tv_nsec
= 0;
824 new_value
.it_interval
.tv_sec
= kvm
->display_time
;
825 new_value
.it_interval
.tv_nsec
= 0;
827 if (timerfd_settime(kvm
->timerfd
, 0, &new_value
, NULL
) != 0) {
828 pr_err("timerfd_settime failed: %d\n", errno
);
838 static int perf_kvm__handle_timerfd(struct perf_kvm_stat
*kvm
)
843 rc
= read(kvm
->timerfd
, &c
, sizeof(uint64_t));
848 pr_err("Failed to read timer fd: %d\n", errno
);
852 if (rc
!= sizeof(uint64_t)) {
853 pr_err("Error reading timer fd - invalid size returned\n");
858 pr_debug("Missed timer beats: %" PRIu64
"\n", c
-1);
865 clear_events_cache_stats(kvm
->kvm_events_cache
);
866 kvm
->total_count
= 0;
868 kvm
->lost_events
= 0;
873 static int fd_set_nonblock(int fd
)
877 arg
= fcntl(fd
, F_GETFL
);
879 pr_err("Failed to get current flags for fd %d\n", fd
);
883 if (fcntl(fd
, F_SETFL
, arg
| O_NONBLOCK
) < 0) {
884 pr_err("Failed to set non-block option on fd %d\n", fd
);
891 static int perf_kvm__handle_stdin(void)
902 static int kvm_events_live_report(struct perf_kvm_stat
*kvm
)
904 int nr_stdin
, ret
, err
= -EINVAL
;
907 /* live flag must be set first */
910 ret
= cpu_isa_config(kvm
);
914 if (!verify_vcpu(kvm
->trace_vcpu
) ||
916 !register_kvm_events_ops(kvm
)) {
920 set_term_quiet_input(&save
);
921 init_kvm_event_record(kvm
);
923 signal(SIGINT
, sig_handler
);
924 signal(SIGTERM
, sig_handler
);
927 if (perf_kvm__timerfd_create(kvm
) < 0) {
932 if (perf_evlist__add_pollfd(kvm
->evlist
, kvm
->timerfd
) < 0)
935 nr_stdin
= perf_evlist__add_pollfd(kvm
->evlist
, fileno(stdin
));
939 if (fd_set_nonblock(fileno(stdin
)) != 0)
942 /* everything is good - enable the events and process */
943 perf_evlist__enable(kvm
->evlist
);
946 struct fdarray
*fda
= &kvm
->evlist
->pollfd
;
949 rc
= perf_kvm__mmap_read(kvm
);
953 err
= perf_kvm__handle_timerfd(kvm
);
957 if (fda
->entries
[nr_stdin
].revents
& POLLIN
)
958 done
= perf_kvm__handle_stdin();
961 err
= fdarray__poll(fda
, 100);
964 perf_evlist__disable(kvm
->evlist
);
972 if (kvm
->timerfd
>= 0)
975 tcsetattr(0, TCSAFLUSH
, &save
);
979 static int kvm_live_open_events(struct perf_kvm_stat
*kvm
)
982 struct perf_evsel
*pos
;
983 struct perf_evlist
*evlist
= kvm
->evlist
;
984 char sbuf
[STRERR_BUFSIZE
];
986 perf_evlist__config(evlist
, &kvm
->opts
, NULL
);
989 * Note: exclude_{guest,host} do not apply here.
990 * This command processes KVM tracepoints from host only
992 evlist__for_each_entry(evlist
, pos
) {
993 struct perf_event_attr
*attr
= &pos
->attr
;
995 /* make sure these *are* set */
996 perf_evsel__set_sample_bit(pos
, TID
);
997 perf_evsel__set_sample_bit(pos
, TIME
);
998 perf_evsel__set_sample_bit(pos
, CPU
);
999 perf_evsel__set_sample_bit(pos
, RAW
);
1000 /* make sure these are *not*; want as small a sample as possible */
1001 perf_evsel__reset_sample_bit(pos
, PERIOD
);
1002 perf_evsel__reset_sample_bit(pos
, IP
);
1003 perf_evsel__reset_sample_bit(pos
, CALLCHAIN
);
1004 perf_evsel__reset_sample_bit(pos
, ADDR
);
1005 perf_evsel__reset_sample_bit(pos
, READ
);
1010 attr
->sample_period
= 1;
1012 attr
->watermark
= 0;
1013 attr
->wakeup_events
= 1000;
1015 /* will enable all once we are ready */
1019 err
= perf_evlist__open(evlist
);
1021 printf("Couldn't create the events: %s\n",
1022 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1026 if (perf_evlist__mmap(evlist
, kvm
->opts
.mmap_pages
, false) < 0) {
1027 ui__error("Failed to mmap the events: %s\n",
1028 str_error_r(errno
, sbuf
, sizeof(sbuf
)));
1029 perf_evlist__close(evlist
);
1040 static int read_events(struct perf_kvm_stat
*kvm
)
1044 struct perf_tool eops
= {
1045 .sample
= process_sample_event
,
1046 .comm
= perf_event__process_comm
,
1047 .ordered_events
= true,
1049 struct perf_data_file file
= {
1050 .path
= kvm
->file_name
,
1051 .mode
= PERF_DATA_MODE_READ
,
1052 .force
= kvm
->force
,
1056 kvm
->session
= perf_session__new(&file
, false, &kvm
->tool
);
1057 if (!kvm
->session
) {
1058 pr_err("Initializing perf session failed\n");
1062 symbol__init(&kvm
->session
->header
.env
);
1064 if (!perf_session__has_traces(kvm
->session
, "kvm record")) {
1070 * Do not use 'isa' recorded in kvm_exit tracepoint since it is not
1071 * traced in the old kernel.
1073 ret
= cpu_isa_config(kvm
);
1077 ret
= perf_session__process_events(kvm
->session
);
1080 perf_session__delete(kvm
->session
);
1084 static int parse_target_str(struct perf_kvm_stat
*kvm
)
1086 if (kvm
->opts
.target
.pid
) {
1087 kvm
->pid_list
= intlist__new(kvm
->opts
.target
.pid
);
1088 if (kvm
->pid_list
== NULL
) {
1089 pr_err("Error parsing process id string\n");
1097 static int kvm_events_report_vcpu(struct perf_kvm_stat
*kvm
)
1100 int vcpu
= kvm
->trace_vcpu
;
1102 if (parse_target_str(kvm
) != 0)
1105 if (!verify_vcpu(vcpu
))
1108 if (!select_key(kvm
))
1111 if (!register_kvm_events_ops(kvm
))
1114 init_kvm_event_record(kvm
);
1117 ret
= read_events(kvm
);
1128 #define STRDUP_FAIL_EXIT(s) \
1136 int __weak
setup_kvm_events_tp(struct perf_kvm_stat
*kvm __maybe_unused
)
1142 kvm_events_record(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1144 unsigned int rec_argc
, i
, j
, events_tp_size
;
1145 const char **rec_argv
;
1146 const char * const record_args
[] = {
1152 const char * const kvm_stat_record_usage
[] = {
1153 "perf kvm stat record [<options>]",
1156 const char * const *events_tp
;
1160 ret
= setup_kvm_events_tp(kvm
);
1162 pr_err("Unable to setup the kvm tracepoints\n");
1166 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++)
1169 rec_argc
= ARRAY_SIZE(record_args
) + argc
+ 2 +
1171 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1173 if (rec_argv
== NULL
)
1176 for (i
= 0; i
< ARRAY_SIZE(record_args
); i
++)
1177 rec_argv
[i
] = STRDUP_FAIL_EXIT(record_args
[i
]);
1179 for (j
= 0; j
< events_tp_size
; j
++) {
1180 rec_argv
[i
++] = "-e";
1181 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm_events_tp
[j
]);
1184 rec_argv
[i
++] = STRDUP_FAIL_EXIT("-o");
1185 rec_argv
[i
++] = STRDUP_FAIL_EXIT(kvm
->file_name
);
1187 for (j
= 1; j
< (unsigned int)argc
; j
++, i
++)
1188 rec_argv
[i
] = argv
[j
];
1190 set_option_flag(record_options
, 'e', "event", PARSE_OPT_HIDDEN
);
1191 set_option_flag(record_options
, 0, "filter", PARSE_OPT_HIDDEN
);
1192 set_option_flag(record_options
, 'R', "raw-samples", PARSE_OPT_HIDDEN
);
1194 set_option_flag(record_options
, 'F', "freq", PARSE_OPT_DISABLED
);
1195 set_option_flag(record_options
, 0, "group", PARSE_OPT_DISABLED
);
1196 set_option_flag(record_options
, 'g', NULL
, PARSE_OPT_DISABLED
);
1197 set_option_flag(record_options
, 0, "call-graph", PARSE_OPT_DISABLED
);
1198 set_option_flag(record_options
, 'd', "data", PARSE_OPT_DISABLED
);
1199 set_option_flag(record_options
, 'T', "timestamp", PARSE_OPT_DISABLED
);
1200 set_option_flag(record_options
, 'P', "period", PARSE_OPT_DISABLED
);
1201 set_option_flag(record_options
, 'n', "no-samples", PARSE_OPT_DISABLED
);
1202 set_option_flag(record_options
, 'N', "no-buildid-cache", PARSE_OPT_DISABLED
);
1203 set_option_flag(record_options
, 'B', "no-buildid", PARSE_OPT_DISABLED
);
1204 set_option_flag(record_options
, 'G', "cgroup", PARSE_OPT_DISABLED
);
1205 set_option_flag(record_options
, 'b', "branch-any", PARSE_OPT_DISABLED
);
1206 set_option_flag(record_options
, 'j', "branch-filter", PARSE_OPT_DISABLED
);
1207 set_option_flag(record_options
, 'W', "weight", PARSE_OPT_DISABLED
);
1208 set_option_flag(record_options
, 0, "transaction", PARSE_OPT_DISABLED
);
1210 record_usage
= kvm_stat_record_usage
;
1211 return cmd_record(i
, rec_argv
, NULL
);
1215 kvm_events_report(struct perf_kvm_stat
*kvm
, int argc
, const char **argv
)
1217 const struct option kvm_events_report_options
[] = {
1218 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1219 "event for reporting: vmexit, "
1220 "mmio (x86 only), ioport (x86 only)"),
1221 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1222 "vcpu id to report"),
1223 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1224 "key for sorting: sample(sort by samples number)"
1225 " time (sort by avg time)"),
1226 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1227 "analyze events only for given process id(s)"),
1228 OPT_BOOLEAN('f', "force", &kvm
->force
, "don't complain, do it"),
1232 const char * const kvm_events_report_usage
[] = {
1233 "perf kvm stat report [<options>]",
1238 argc
= parse_options(argc
, argv
,
1239 kvm_events_report_options
,
1240 kvm_events_report_usage
, 0);
1242 usage_with_options(kvm_events_report_usage
,
1243 kvm_events_report_options
);
1246 if (!kvm
->opts
.target
.pid
)
1247 kvm
->opts
.target
.system_wide
= true;
1249 return kvm_events_report_vcpu(kvm
);
1252 #ifdef HAVE_TIMERFD_SUPPORT
1253 static struct perf_evlist
*kvm_live_event_list(void)
1255 struct perf_evlist
*evlist
;
1256 char *tp
, *name
, *sys
;
1258 const char * const *events_tp
;
1260 evlist
= perf_evlist__new();
1264 for (events_tp
= kvm_events_tp
; *events_tp
; events_tp
++) {
1266 tp
= strdup(*events_tp
);
1270 /* split tracepoint into subsystem and name */
1272 name
= strchr(tp
, ':');
1274 pr_err("Error parsing %s tracepoint: subsystem delimiter not found\n",
1282 if (perf_evlist__add_newtp(evlist
, sys
, name
, NULL
)) {
1283 pr_err("Failed to add %s tracepoint to the list\n", *events_tp
);
1295 perf_evlist__delete(evlist
);
1302 static int kvm_events_live(struct perf_kvm_stat
*kvm
,
1303 int argc
, const char **argv
)
1305 char errbuf
[BUFSIZ
];
1308 const struct option live_options
[] = {
1309 OPT_STRING('p', "pid", &kvm
->opts
.target
.pid
, "pid",
1310 "record events on existing process id"),
1311 OPT_CALLBACK('m', "mmap-pages", &kvm
->opts
.mmap_pages
, "pages",
1312 "number of mmap data pages",
1313 perf_evlist__parse_mmap_pages
),
1314 OPT_INCR('v', "verbose", &verbose
,
1315 "be more verbose (show counter open errors, etc)"),
1316 OPT_BOOLEAN('a', "all-cpus", &kvm
->opts
.target
.system_wide
,
1317 "system-wide collection from all CPUs"),
1318 OPT_UINTEGER('d', "display", &kvm
->display_time
,
1319 "time in seconds between display updates"),
1320 OPT_STRING(0, "event", &kvm
->report_event
, "report event",
1321 "event for reporting: "
1322 "vmexit, mmio (x86 only), ioport (x86 only)"),
1323 OPT_INTEGER(0, "vcpu", &kvm
->trace_vcpu
,
1324 "vcpu id to report"),
1325 OPT_STRING('k', "key", &kvm
->sort_key
, "sort-key",
1326 "key for sorting: sample(sort by samples number)"
1327 " time (sort by avg time)"),
1328 OPT_U64(0, "duration", &kvm
->duration
,
1329 "show events other than"
1330 " HLT (x86 only) or Wait state (s390 only)"
1331 " that take longer than duration usecs"),
1332 OPT_UINTEGER(0, "proc-map-timeout", &kvm
->opts
.proc_map_timeout
,
1333 "per thread proc mmap processing timeout in ms"),
1336 const char * const live_usage
[] = {
1337 "perf kvm stat live [<options>]",
1340 struct perf_data_file file
= {
1341 .mode
= PERF_DATA_MODE_WRITE
,
1345 /* event handling */
1346 kvm
->tool
.sample
= process_sample_event
;
1347 kvm
->tool
.comm
= perf_event__process_comm
;
1348 kvm
->tool
.exit
= perf_event__process_exit
;
1349 kvm
->tool
.fork
= perf_event__process_fork
;
1350 kvm
->tool
.lost
= process_lost_event
;
1351 kvm
->tool
.ordered_events
= true;
1352 perf_tool__fill_defaults(&kvm
->tool
);
1355 kvm
->display_time
= 1;
1356 kvm
->opts
.user_interval
= 1;
1357 kvm
->opts
.mmap_pages
= 512;
1358 kvm
->opts
.target
.uses_mmap
= false;
1359 kvm
->opts
.target
.uid_str
= NULL
;
1360 kvm
->opts
.target
.uid
= UINT_MAX
;
1361 kvm
->opts
.proc_map_timeout
= 500;
1364 disable_buildid_cache();
1369 argc
= parse_options(argc
, argv
, live_options
,
1372 usage_with_options(live_usage
, live_options
);
1375 kvm
->duration
*= NSEC_PER_USEC
; /* convert usec to nsec */
1378 * target related setups
1380 err
= target__validate(&kvm
->opts
.target
);
1382 target__strerror(&kvm
->opts
.target
, err
, errbuf
, BUFSIZ
);
1383 ui__warning("%s", errbuf
);
1386 if (target__none(&kvm
->opts
.target
))
1387 kvm
->opts
.target
.system_wide
= true;
1391 * generate the event list
1393 err
= setup_kvm_events_tp(kvm
);
1395 pr_err("Unable to setup the kvm tracepoints\n");
1399 kvm
->evlist
= kvm_live_event_list();
1400 if (kvm
->evlist
== NULL
) {
1405 symbol_conf
.nr_events
= kvm
->evlist
->nr_entries
;
1407 if (perf_evlist__create_maps(kvm
->evlist
, &kvm
->opts
.target
) < 0)
1408 usage_with_options(live_usage
, live_options
);
1413 kvm
->session
= perf_session__new(&file
, false, &kvm
->tool
);
1414 if (kvm
->session
== NULL
) {
1418 kvm
->session
->evlist
= kvm
->evlist
;
1419 perf_session__set_id_hdr_size(kvm
->session
);
1420 ordered_events__set_copy_on_queue(&kvm
->session
->ordered_events
, true);
1421 machine__synthesize_threads(&kvm
->session
->machines
.host
, &kvm
->opts
.target
,
1422 kvm
->evlist
->threads
, false, kvm
->opts
.proc_map_timeout
);
1423 err
= kvm_live_open_events(kvm
);
1427 err
= kvm_events_live_report(kvm
);
1430 perf_session__delete(kvm
->session
);
1431 kvm
->session
= NULL
;
1432 perf_evlist__delete(kvm
->evlist
);
1438 static void print_kvm_stat_usage(void)
1440 printf("Usage: perf kvm stat <command>\n\n");
1442 printf("# Available commands:\n");
1443 printf("\trecord: record kvm events\n");
1444 printf("\treport: report statistical data of kvm events\n");
1445 printf("\tlive: live reporting of statistical data of kvm events\n");
1447 printf("\nOtherwise, it is the alias of 'perf stat':\n");
1450 static int kvm_cmd_stat(const char *file_name
, int argc
, const char **argv
)
1452 struct perf_kvm_stat kvm
= {
1453 .file_name
= file_name
,
1456 .report_event
= "vmexit",
1457 .sort_key
= "sample",
1462 print_kvm_stat_usage();
1466 if (!strncmp(argv
[1], "rec", 3))
1467 return kvm_events_record(&kvm
, argc
- 1, argv
+ 1);
1469 if (!strncmp(argv
[1], "rep", 3))
1470 return kvm_events_report(&kvm
, argc
- 1 , argv
+ 1);
1472 #ifdef HAVE_TIMERFD_SUPPORT
1473 if (!strncmp(argv
[1], "live", 4))
1474 return kvm_events_live(&kvm
, argc
- 1 , argv
+ 1);
1478 return cmd_stat(argc
, argv
, NULL
);
1480 #endif /* HAVE_KVM_STAT_SUPPORT */
1482 static int __cmd_record(const char *file_name
, int argc
, const char **argv
)
1484 int rec_argc
, i
= 0, j
;
1485 const char **rec_argv
;
1487 rec_argc
= argc
+ 2;
1488 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1489 rec_argv
[i
++] = strdup("record");
1490 rec_argv
[i
++] = strdup("-o");
1491 rec_argv
[i
++] = strdup(file_name
);
1492 for (j
= 1; j
< argc
; j
++, i
++)
1493 rec_argv
[i
] = argv
[j
];
1495 BUG_ON(i
!= rec_argc
);
1497 return cmd_record(i
, rec_argv
, NULL
);
1500 static int __cmd_report(const char *file_name
, int argc
, const char **argv
)
1502 int rec_argc
, i
= 0, j
;
1503 const char **rec_argv
;
1505 rec_argc
= argc
+ 2;
1506 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1507 rec_argv
[i
++] = strdup("report");
1508 rec_argv
[i
++] = strdup("-i");
1509 rec_argv
[i
++] = strdup(file_name
);
1510 for (j
= 1; j
< argc
; j
++, i
++)
1511 rec_argv
[i
] = argv
[j
];
1513 BUG_ON(i
!= rec_argc
);
1515 return cmd_report(i
, rec_argv
, NULL
);
1519 __cmd_buildid_list(const char *file_name
, int argc
, const char **argv
)
1521 int rec_argc
, i
= 0, j
;
1522 const char **rec_argv
;
1524 rec_argc
= argc
+ 2;
1525 rec_argv
= calloc(rec_argc
+ 1, sizeof(char *));
1526 rec_argv
[i
++] = strdup("buildid-list");
1527 rec_argv
[i
++] = strdup("-i");
1528 rec_argv
[i
++] = strdup(file_name
);
1529 for (j
= 1; j
< argc
; j
++, i
++)
1530 rec_argv
[i
] = argv
[j
];
1532 BUG_ON(i
!= rec_argc
);
1534 return cmd_buildid_list(i
, rec_argv
, NULL
);
1537 int cmd_kvm(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1539 const char *file_name
= NULL
;
1540 const struct option kvm_options
[] = {
1541 OPT_STRING('i', "input", &file_name
, "file",
1543 OPT_STRING('o', "output", &file_name
, "file",
1544 "Output file name"),
1545 OPT_BOOLEAN(0, "guest", &perf_guest
,
1546 "Collect guest os data"),
1547 OPT_BOOLEAN(0, "host", &perf_host
,
1548 "Collect host os data"),
1549 OPT_STRING(0, "guestmount", &symbol_conf
.guestmount
, "directory",
1550 "guest mount directory under which every guest os"
1551 " instance has a subdir"),
1552 OPT_STRING(0, "guestvmlinux", &symbol_conf
.default_guest_vmlinux_name
,
1553 "file", "file saving guest os vmlinux"),
1554 OPT_STRING(0, "guestkallsyms", &symbol_conf
.default_guest_kallsyms
,
1555 "file", "file saving guest os /proc/kallsyms"),
1556 OPT_STRING(0, "guestmodules", &symbol_conf
.default_guest_modules
,
1557 "file", "file saving guest os /proc/modules"),
1558 OPT_INCR('v', "verbose", &verbose
,
1559 "be more verbose (show counter open errors, etc)"),
1563 const char *const kvm_subcommands
[] = { "top", "record", "report", "diff",
1564 "buildid-list", "stat", NULL
};
1565 const char *kvm_usage
[] = { NULL
, NULL
};
1570 argc
= parse_options_subcommand(argc
, argv
, kvm_options
, kvm_subcommands
, kvm_usage
,
1571 PARSE_OPT_STOP_AT_NON_OPTION
);
1573 usage_with_options(kvm_usage
, kvm_options
);
1579 file_name
= get_filename_for_perf_kvm();
1582 pr_err("Failed to allocate memory for filename\n");
1587 if (!strncmp(argv
[0], "rec", 3))
1588 return __cmd_record(file_name
, argc
, argv
);
1589 else if (!strncmp(argv
[0], "rep", 3))
1590 return __cmd_report(file_name
, argc
, argv
);
1591 else if (!strncmp(argv
[0], "diff", 4))
1592 return cmd_diff(argc
, argv
, NULL
);
1593 else if (!strncmp(argv
[0], "top", 3))
1594 return cmd_top(argc
, argv
, NULL
);
1595 else if (!strncmp(argv
[0], "buildid-list", 12))
1596 return __cmd_buildid_list(file_name
, argc
, argv
);
1597 #ifdef HAVE_KVM_STAT_SUPPORT
1598 else if (!strncmp(argv
[0], "stat", 4))
1599 return kvm_cmd_stat(file_name
, argc
, argv
);
1602 usage_with_options(kvm_usage
, kvm_options
);